detecting tamarisk defoliation, the forest disturbance index
(Healey
et al
., 2005) and a decision-tree model (random
forests) (Breiman, 2001). Nagler
et al
. (2012) developed an
approach for estimating regional evapotranspiration (
ET
)
and foliage density changes caused by beetles using Landsat
TM
and
MODIS
data. Their results for six western US rivers
indicated that defoliation events contributed to about 15
percent of the overall
ET
and foliage density reduction, with
marked variations among river systems. Snyder
et al
. (2012)
compared
ET
and carbon flux measured by eddy covariance
to
NDVI
calculated from Landsat
TM
and
ETM
+ data. Declines
in
NDVI
occurred during summer decreases in leaf area and
ET
caused by defoliation (Snyder
et al
., 2012). Nagler
et al
. (2014)
synthesized
MODIS
data, networked digital camera images and
ground surveys to track beetle dispersal and its impacts on the
Virgin River from 2010 to 2013. They concluded that beetle
damage progressed at a rate of about 25 km yr
-1
, much faster
than previous expectations, and caused a 50 percent reduc-
tion in leaf area index and
ET
of tamarisk stands by 2012.
Identification of desiccated or dead tamarisk canopies may
be aided by the availability of high spatial resolution (e.g.,
GeoEye, WorldView) or hyperspectral (e.g., Hyperion,
AVIRIS
)
remote sensing data. High spatial resolution images may re-
duce spectral mixing at the stand scale to allow the separation
of desiccated and dead canopies (Dennison
et al.
2009, Meng
et al.
2012), and can estimate
ET
at plant canopy scales (Nouri,
2014). In contrast with multispectral data, spectroscopic (hy-
perspectral) analysis can resolve spectral features related to
vegetation structure and biochemistry using hundreds of near-
contiguous narrow bandwidth channels (Ustin
et al
., 2004).
As a result, hyperspectral data have been used to assess spec-
tral separability among different vegetation species (Dennison
and Roberts, 2003; Pu, 2009; Van Aardt and Wynne, 2001),
to estimate the change in biochemical compounds caused
by disturbance or stress (Bian
et al
., 2010; Estep and Carter,
2005; Pu
et al
., 2008), and to distinguish between green plant,
plant litter, and soil at both leaf and canopy levels (Nagler
et
al
., 2000; Nagler
et al
., 2003; Inoue
et al
., 2008). After measur-
ing spectral reflectance of plant litter and soil samples using
a spectroradiometer, Nagler
et al
. (2000 and 2003) found
no unique spectral feature for discrimination of plant litter
and soil existed in the visible or near infrared (
NIR
) spectral
regions. However, in the shortwave infrared (
SWIR
) region, an
absorption feature associated with cellulose and lignin was
found and a corresponding spectral index called cellulose
absorption index (
CAI
) was designed to quantify plant litter
cover (Nagler
et al
., 2000; Nagler
et al
., 2003).
Various non-photosynthetic vegetation and green vegeta-
tion cover types have been successfully classified through
spectral matching and linear mixture modeling techniques
(Cochrane, 1998; Roberts
et al
., 1998; Datt, 2000; Hostert
et
al
., 2003; Herold
et al
., 2004; Daigo
et al
., 2004; Zhang
et al
.,
2006; Sonnentag
et al
., 2007; Zhang
et al
., 2007; Pacheco and
McNairn, 2010; Haest
et al
., 2013; Somers and Asner, 2013).
Spectral matching is one of the most widely used spectroscop-
ic techniques, aiming to detect targeted pixels or endmembers,
while linear spectral mixture analysis (
LSMA
) is designed for
disaggregating mixed spectral pixels from remote sensing data
sets. Asner and Lobell (2000) claimed that a careful selection
of wavelengths or spectral features for
LSMA
may improve
classification accuracy and reduce computation complexity.
Somers
et al
. (2010) developed and tested an automated
LSMA
algorithm, known as stable zone unmixing (
SZU
), to overcome
the limitations of the
AutoSWIR
algorithm presented by Asner
and Lobell (2000). The instability index (
ISI
) was calculated to
select stable spectral features (Somers
et al
., 2008).
ISI
account-
ed for both the spectral variability within a class and the spec-
tral similarity among classes to indicate the most useful and
stable wavelength ranges over the full spectral range. Somers
and Asner (2013) demonstrated that a proper wavelength
selection strategy could avoid redundant information and
improve classi cation accuracies, by emphasizing the subtle
spectral and phenological differences among targeted classes.
Discrimination of defoliated and dead tamarisk canopy types
could benefit from a similar wavelength selection strategy.
Material and Methods
Study Site and Spectral Measurement
The study site was located at the University of Utah Rio Mesa
Center, 65 km northeast of Moab in southeastern Utah. The
riparian corridors along the Dolores River at Rio Mesa Center
consist of dense tamarisk stands as well as some native cot-
tonwood (
Populus fremontii
) and willow (
Salix gooddingii
)
trees. As a first step towards spectroscopic analysis, the re-
flectance of green, desiccated and dead tamarisk canopies was
measured
in situ
along the Dolores River over the 350-2,500
nm wavelength range using an Analytical Spectral Devices
(
ASD
) field spectrometer with a 25° field of view (PANalyti-
cal; Analytic Spectral Devices, Boulder, Colorado). A white
spectralon standard was used to calibrate the spectral mea-
surements. The full-width-at-half maximum (
FWHM
) and the
sampling interval of the spectrometer for the 350-1,050 nm
spectral range were 3 nm and 1.4 nm, respectively. Over the
1,050-2,500 nm spectral range, the
FWHM
and sampling inter-
val were 8 nm and 2 nm. A Gaussian function with a 5 nm
FWHM
was used to resample the 1 nm instrument output.
The reflectance measurements were carried out in early
October 2013 within two hours before or after solar noon to
reduce solar zenith angle effects, and under cloudless sky
conditions. Reflectance spectra were measured from nadir at a
height of approximately 15 cm above the canopy using a pis-
tol grip with extension. The representative tamarisk canopy
types (green, desiccated and dead) with different heights
along riparian corridors were selected
in situ
to ensure signifi-
cant variations in ground cover fractions and reflectance spec-
tra. Desiccated tamarisk canopies showed two colors: yellow
and brown (Plate 1). Dead canopies were devoid of desiccated
leaf material, with only branches showing (Plate 1). Reflec-
tance measurements for selected targets were performed five
times and averaged for analysis. In total, 67 canopy spectra
were collected and analyzed including 17 green, 15 dead, 27
yellow desiccated, and 8 brown desiccated canopies. Means
and standard deviations for reflectance spectra of green,
brown desiccated, yellow desiccated and dead canopies are
shown in Figure 1. The major absorption regions influenced
by atmospheric water vapor content were excluded from
analysis (Somers
et al
., 2009).
Cross Correlogram Spectral Matching
A suitable number of mapping classes is important for im-
proving classification accuracy and efficiency (Richards and
Jia, 2006). Desiccated tamarisk canopies showed two distinct
colors (yellow and brown)
in situ
(Plate 1) representing dif-
ferent desiccated status, while Figure 1 indicated the spectral
signatures of yellow and brown canopies were similar. Con-
sidering the computational task of remote sensing classifica-
tion, it may be beneficial to examine if spectral separability
between yellow and brown desiccated canopies is minor or
if it is reasonable to split the desiccated canopy type into two
subtypes (yellow and brown). A spectral matching tech-
nique called cross correlogram spectral matching (
CCSM
) was
implemented to compare the spectral separability of different
tamarisk canopy types (Van Der Meer and Bakker, 1997).
CCSM
compares the differences between a reference spectrum and
an unknown spectrum in the form of reference amplitude as
200
March 2015
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
